Bidirectional Support: Unlike the previous version which only supported En -> TlH, this model handles both directions seamlessly.
Fixed Stuttering: The " 'oH, 'oH! " repetition issue has been resolved by doubling the training data and shuffling direction tags.
Zero-Shot Potential: Due to the multilingual nature of M2M100, it can perform zero-shot translation from Romanian (ro) to Klingon with decent accuracy.

📊 Training Details

Dataset: 66,176 rows (33k En->TlH + 33k TlH->En mirrored).
Epochs: 3
Batch Size: 16 (4 per device + 4 gradient accumulation steps).
Final Training Loss: ~0.72

🛠️ Usage

python

from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer

model = M2M100ForConditionalGeneration.from_pretrained("MihaiPopa-1/M2M100-418M-Klingon")
tokenizer = M2M100Tokenizer.from_pretrained("MihaiPopa-1/M2M100-418M-Klingon")

def translate(text, src="en", tgt="kn"):
    tokenizer.src_lang = src
    encoded = tokenizer(text, return_tensors="pt")
    generated_tokens = model.generate(**encoded, forced_bos_token_id=tokenizer.get_lang_id(tgt))
    return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]

# Examples:
print(translate("Victory is ours!", src="en", tgt="kn")) # English to Klingon
print(translate("qapla'!", src="kn", tgt="en"))          # Klingon to English
print(translate("Războiul a început.", src="ro", tgt="kn")) # Romanian to Klingon (Zero-shot)